CAMEL

CAMEL (Communicative Agents for Machine Learning) is a multi-agent framework that prompts two language models to cooperate on a task by assigning them complementary roles — a human surrogate that writes instruction-following data and an assistant that learns from it — without relying on human-curated datasets.

The human surrogate proposes a task goal while the assistant asks clarifying questions and works through the problem step by step, with the human surrogate providing feedback until a satisfactory solution is reached. This role-playing dynamic produces a natural curriculum of increasing complexity, starting from simple tasks and progressing to multi-step reasoning that neither agent would produce alone. CAMEL demonstrated that two models talking to each other can bootstrap a full instruction dataset that rivals human-written data in quality.

Role assignment and initial prompts heavily influence the resulting data quality. The approach can occasionally produce nonsensical or misaligned instructions that require filtering. CAMEL works best when the initial prompts constrain the agents toward constructive rather than adversarial behavior, which requires careful prompt engineering. The approach opens the door to self-improving AI systems that generate their own training signal.

Whether emergent role dynamics produce higher quality data than structured human-designed prompts is unclear. How to prevent the agents from colluding to produce convincing but incorrect or biased instructions is an open problem. The gap between CAMEL-generated data and genuinely human-originated instructions in terms of measurable properties is not well-characterized.